Cooling as a constraint for AI

Electricity that flows into a GPU does not disappear. Almost all of it becomes heat. A single modern accelerator can produce as much heat as a space heater, and a full rack packs dozens of them into a small footprint. If that heat is not removed quickly, the chips throttle to protect themselves, and at the extreme they fail outright.

This makes cooling a genuine constraint, not a detail. A site can have chips and power available and still be unable to run them at full speed, simply because it cannot carry the heat away fast enough. In that situation, the cooling system, not the chip, is what caps performance.

Heat also compounds with density. The more compute you pack into a rack to use space efficiently, the more concentrated the heat becomes, which makes the cooling challenge harder exactly where you most want to push performance.

For years, data centers cooled servers by moving large volumes of cold air. As AI racks grew denser and hotter, air struggled to keep up. Moving enough air to cool a high-density AI rack becomes loud, inefficient, and eventually impossible past a certain heat level, because air simply cannot carry heat away fast enough.

That is why the industry is shifting toward liquid cooling, which carries heat away far more effectively than air. Liquid cooling lets operators pack more compute into the same space, but it also adds plumbing, design, and operational complexity that has to be managed carefully. A leak or a pump failure becomes a serious event, so reliability engineering matters more than ever.

There are several approaches along the spectrum. Some facilities bring liquid directly to a cold plate on each chip, while others immerse whole servers in a non-conductive fluid. Each method trades cost, complexity, and serviceability differently, but they share the same goal: move heat away from dense hardware faster than air ever could, so the chips can keep running at full speed.

Because heat sets a ceiling, cooling capacity directly shapes how much AI a facility can run. The IEA reports data centres used about 415 TWh worldwide in 2024 and could reach around 945 TWh by 2030, and cooling is part of every one of those figures. Better cooling means more of the energy goes to useful compute instead of being wasted moving heat around.

This is why cooling sits alongside chips and power as one of the core constraints on AI. A well-cooled facility can safely run more hardware, more of the time, which is the whole point of owning the hardware in the first place.

It also helps explain why two sites with identical chips can deliver very different results. The one with better cooling can sustain higher performance without throttling, fit more hardware into its floor space, and keep components within safe limits for longer. Cooling, in other words, is not a background utility. It is part of what determines how much value the hardware can produce.

The heat a data center removes does not have to be pure waste. Some facilities capture it and put it to use, warming nearby buildings or feeding district heating systems. Reusing heat does not lower the amount a chip produces, but it can turn a disposal problem into a modest benefit and improve the overall efficiency of a site.

Heat reuse works best where there is something nearby that needs warmth and a climate that makes the economics sensible. It is not available everywhere, and it adds its own engineering, so it remains the exception rather than the rule. Still, it points to a future where cooling is designed as part of a larger energy system rather than treated only as a cost.

Whether or not heat is reused, the core constraint stands. The faster and more completely a facility can move heat away from dense hardware, the more compute it can run safely, which keeps cooling at the center of how much AI a site can deliver.